Deep learning using remote sensing to predict the occurrence of soil water repellency in dryland pastures in New Zealand

Abstract

Soil water repellency (SWR) is natural phenomenon occurring in soils throughout the world. This property is related to the types of organic matter present in soils. Some of the organic matter can be classified as hydrophilic, attracted to water; while others are hydrophobic, repelled by water. The inability of soils to absorb water after an irrigation or a rainfall event, increases the potential for runoff and nutrients losses to waterways. Its impact on ecosystems services is influenced at multiple temporal and spatial scales. However, the mechanisms of how and why SWR develops is still poorly understood. Thus, there is no current method that allows for the prediction of soil SWR occurrence. In dryland pasture, SWR development is controlled by the cycling of hydrophobic materials. Detailed spatiotemporal description of these systems would allow better understanding of this phenomenon. However, acquiring data that reflects perfectly the variability in these systems is difficult. Cycling of organic matter and hydrophobic materials is generally controlled by land use, management, climate, and soil properties. In addition to climate and soil, the fate of these materials in pastoral systems is controlled by diverse drivers (e.g. grazing management, species, nutrients input). This complex interaction between soils, climate and management; makes it a difficult system to understand and model. Deep learning of the time series remote sensing data shows promise to predict the occurrence of SWR in dryland pastures. Multispectral and synthetic aperture radar data are highly correlated with land cover and surface soil moisture states, respectively. The time series data is a good reflection of how the moisture dynamics changes under the land cover, which is indicative of the expression of SWR. The aim of this study is to investigate ability of deep learning on remote sensed data to predict occurrence of SWR in a New Zealand (NZ) grassland system. The result from this study shows that deep learning classification methods, including artificial neural networks, random forest and support vector machine; can accurately predict the levels of potential soil water repellency in dryland pastures. With improved prediction of SWR occurrence in NZ pastoral systems, better management of fertilizer and irrigation application can be adopted.